Abstract: During fish farming, the number of fish in the ponds needs to be monitored regularly. Aiming at the problem of missing detection in the existing methods, this paper proposed a local optimization method based on the YOLOv5s model and a fish quantity statistics method with an improved output scale. By adding local information such as the head and tail of the detected fish, the category with the largest number was selected from the three categories of the whole body, head, and tail of the fish as the result of quantitative statistics to solve the problem of missing detection. At the same time, for the case that the whole body, head, and tail of fish were displayed as large-scale or mesoscale targets in the image, the feature output of these two types of targets was increased to improve the target detection ability of the model, so that the model can be suitable for quantitative detection under the current conditions. The results showed that compared with manual counting, the error of the quantity counted by this method was small, the accuracy was 96.3%, and the detection frame rate was 111 FPS. Based on the YOLOv5 model, the application of local optimization strategy increased the number of statistics by 37.4%, and the improvement of output scale increases the number of statistics by 4.9%. The study can be applied to the statistics of fish stocks in fishery and fish detection.

Key words: fish quantity statistics, object detection, local optimization, improved output scale

摘要： 在渔业养殖过程中，需要定期对养殖池内鱼的数量进行监测。针对现有方法中存在的漏检问题，提出基于YOLOv5模型的局部优选以及改进输出尺度的鱼类数量统计方法。通过增加检测鱼的头部、尾部等局部信息，从鱼的全身、鱼的头部、鱼的尾部三个类别中优选数量最多的类作为数量统计的结果以解决漏检的问题。同时，针对鱼的全身、鱼的头部和鱼的尾部在图像中显示为大尺度或中尺度目标的情况，增加了这两类目标的特征输出以提高模型对目标的检测能力，使得模型能够适用于当前条件下的数量检测。结果显示，通过本方法统计出的数量与人工计数相比误差较小，准确率为96.3%，检测的帧速率为111 fps。在YOLOv5模型的基础上，应用局部优选的策略使得统计数量提高了37.4%，对输出尺度的改进使得统计数量提高了4.9%。该研究可以应用于渔业养殖鱼群数量统计和鱼类检测等方面。

关键词: 鱼类数量统计, 目标检测, 局部优选, 改进输出尺度